2865 Security Features 9a7d26a9

MCP Gateway Security Features

Current Version: 1.0.0-RC-1 — The gateway ships with the controls described below. Everything listed here is present in the codebase today; future roadmap items live in docs/docs/architecture/roadmap.md.

Security Posture Overview

• Authentication on by default. AUTH_REQUIRED=true ensures every API, SSE, and Admin UI route requires an authenticated session unless you explicitly opt out for local testing.
• Secure defaults for operators. The service binds to 127.0.0.1 unless overridden, MCPGATEWAY_UI_ENABLED=false and MCPGATEWAY_ADMIN_API_ENABLED=false keep the Admin UI/API offline in production, and cookies are hardened with SECURE_COOKIES=true, HttpOnly, and SameSite=lax.
• Environment-aware CORS & cookies. ENVIRONMENT and APP_DOMAIN drive cors/origin policy, switching between a localhost allowlist in development and strict origin checks in production (mcpgateway/middleware/security_headers.py).
• Security posture reporting. Startup invokes validate_security_configuration() (mcpgateway/main.py) which consumes settings.get_security_status() to log weak secrets, missing auth, or insecure federation setups. Setting REQUIRE_STRONG_SECRETS=true upgrades warnings to hard failures.

Authentication & Identity

Core Gateway Authentication

• HTTP Basic Auth is disabled by default for security. Enable with API_ALLOW_BASIC_AUTH=true for API endpoints or DOCS_ALLOW_BASIC_AUTH=true for docs. When enabled, credentials use BASIC_AUTH_USER/BASIC_AUTH_PASSWORD. The Admin UI uses email/password authentication, not Basic auth.
• JWT bearer tokens are required for API access and MCP transports when MCP_CLIENT_AUTH_ENABLED=true (default). For reverse proxies you can opt into TRUST_PROXY_AUTH=true and provide the authenticated identity through PROXY_USER_HEADER.
• Token issuance tooling. python -m mcpgateway.utils.create_jwt_token produces gateway-signed tokens for automation. The helper respects configured expiry, issuer, and audience claims.

JWT Token Management

• Algorithm agility. mcpgateway/utils/jwt_config_helper.py supports HS256/384/512, RS256/384/512, and ES256/384/512. For asymmetric algorithms the helper validates key paths on startup and reads the PEM material securely.
• Secret validation. The Pydantic field validator (Settings.validate_secrets) logs warnings for default or low-entropy secrets, and when REQUIRE_STRONG_SECRETS=true startup fails if critical values remain weak.
• Revocation and audit. API tokens are modelled as JWTs with per-token jti identifiers. Revocations (TokenRevocation) and usage logs (TokenUsageLog) persist to the database, enabling immediate invalidation and monitoring.

JWT ID (JTI) Claim

The jti (JWT ID) claim is a unique identifier for each JWT token, defined in RFC 7519 Section 4.1.7. MCP Gateway uses JTI for:

1. Token Revocation: Each token can be individually revoked by its JTI without invalidating all tokens for a user. The TokenRevocation table stores revoked JTIs.
2. Auth Cache Keying: The authentication cache uses {email}:{jti} as the cache key pattern (mcpgateway/cache/auth_cache.py). This enables per-token caching and prevents cache collisions when users have multiple active tokens.
3. Replay Attack Prevention: JTIs enable detection of token reuse, allowing the gateway to track and limit how many times a specific token is used.
4. Audit Trails: Every TokenUsageLog entry records the JTI, enabling detailed per-token usage analytics and anomaly detection.

Token Generation Examples:

# Email auth tokens (always include JTI)
# Location: mcpgateway/routers/email_auth.py
payload = {
    "sub": user.email,
    "jti": str(uuid.uuid4()),  # Unique per token
    ...
}

# Load test tokens (configurable)
# Location: tests/loadtest/locustfile.py
payload = {
    "sub": JWT_USERNAME,
    "jti": str(uuid.uuid4()),  # Added for proper cache keying
    "exp": datetime.now(timezone.utc) + timedelta(hours=JWT_TOKEN_EXPIRY_HOURS),
    ...
}

Cache Behavior:

• Tokens with JTI: Cache key is mcpgw:auth:ctx:{email}:{jti-uuid}
• Tokens without JTI: Cache key is mcpgw:auth:ctx:{email}:no-jti

For production deployments, always include JTI in issued tokens to enable proper caching, revocation, and audit capabilities.

Email-Based Authentication

• Argon2id password hashing. EmailAuthService hashes credentials with configurable ARGON2ID_TIME_COST, ARGON2ID_MEMORY_COST, and ARGON2ID_PARALLELISM.
• Password policy controls. Minimum length and complexity are driven by PASSWORD_MIN_LENGTH and related flags. Attempts that fail policy raise PasswordValidationError.
• Account lockout & auditing. MAX_FAILED_LOGIN_ATTEMPTS and ACCOUNT_LOCKOUT_DURATION_MINUTES enforce lockouts at the ORM layer (EmailUser.increment_failed_attempts). Every login, registration, or password change emits an EmailAuthEvent with IP and user-agent metadata for investigation.
• Admin bootstrap. The first superuser is provisioned from PLATFORM_ADMIN_EMAIL/PLATFORM_ADMIN_PASSWORD, enabling secure initial access even when SSO is not configured.

Token Catalog & API Keys

• Hashed personal and team tokens. TokenCatalogService stores only SHA-256 hashes (token_hash) of issued tokens (EmailApiToken) and keeps the raw secret in memory just long enough to present it once.
• Fine-grained scopes. Tokens can be confined to a server, a permission list, IP CIDRs, and time-of-day or usage quotas via TokenScope.
• Usage analytics. Every request records method, endpoint, IP, user-agent, latency, and block reason in TokenUsageLog, supporting anomaly detection.
• Immediate revocation. TokenRevocation entries are enforced before gateway execution, guaranteeing revoked tokens cannot be replayed.

OAuth & SSO Federation

• Multi-provider SSO. mcpgateway/services/sso_service.py supports GitHub, Google, IBM Security Verify, Microsoft Entra ID, Okta, Keycloak, and generic OIDC providers. Secrets are encrypted with a Fernet key derived from AUTH_ENCRYPTION_SECRET.
• Security-state tracking. SSOAuthSession persists OAuth state tokens, PKCE code_verifier, and nonces to prevent CSRF and replay attacks.
• Per-user OAuth vault. TokenStorageService encrypts access/refresh tokens using AES-GCM (oauth_encryption.py) and keys them by both gateway and gateway user to prevent cross-tenant leakage.
• Dynamic Client Registration (DCR). DcrService discovers OAuth metadata (RFC 8414), honours issuer allowlists (DCR_ALLOWED_ISSUERS), registers clients, and encrypts the resulting client secrets and registration access tokens before storing them (RegisteredOAuthClient).
• Tool credential encryption. The same AUTH_ENCRYPTION_SECRET powers services_auth.encode_auth() to store upstream tool auth blobs as AES-GCM tokens inside the database.

Authorization & Access Control

• Role-Based Access Control (RBAC). PermissionService and RoleService implement global/team/personal scopes with caching, inheritance, and audit logging (PermissionAuditLog). Admin bypass is explicit, and permission checks default to deny on error.
• Multi-tenancy primitives. Teams, invites, and memberships (EmailTeam, EmailTeamMember, TeamInvitationService) enforce owner-only invitations, configurable expiry, and per-team quotas (MAX_TEAMS_PER_USER, MAX_MEMBERS_PER_TEAM). Personal teams can be auto-created with AUTO_CREATE_PERSONAL_TEAMS=true.
• Resource visibility. Tools, prompts, resources, and gateways include a visibility flag (private/team/public) that PermissionService respects when resolving access.
• Feature gating. Administrative capabilities stay off unless you opt in: MCPGATEWAY_UI_ENABLED, MCPGATEWAY_ADMIN_API_ENABLED, MCPGATEWAY_BULK_IMPORT_ENABLED, MCPGATEWAY_CATALOG_ENABLED, and MCPGATEWAY_A2A_ENABLED all default to safe values.
• Scoped API credentials. Tokens can be restricted to individual virtual servers, explicit permission strings, and IP ranges; blocked requests are captured via TokenUsageLog.blocked.
• Header passthrough controls. utils/passthrough_headers.py keeps passthrough disabled unless ENABLE_HEADER_PASSTHROUGH=true, sanitises header names/values, rejects conflicting Authorization headers, and lets clients safely supply X-Upstream-Authorization for upstream delegation.
• Policy-as-code plugins. The plugin framework powers deny/allow decisions before and after prompt/tool/resource execution. Security-focused plugins include deny_filter, pii_filter, content_moderation, output_length_guard, schema_guard, sql_sanitizer, secrets_detection, rate_limiter, url_reputation, vault, watchdog, and the optional external OPA integration for Rego policies (plugins/external/opa).

Data Protection & Secret Handling

• AES-GCM secret vault. mcpgateway/utils/services_auth.py derives a 32-byte key from AUTH_ENCRYPTION_SECRET and encrypts tool/resource credentials, ensuring secrets stored in the database or logs are opaque without the passphrase.
• Encrypted OAuth/SSO secrets. The SSO service and DCR service wrap client secrets and registration tokens with Fernet and only decrypt them on demand.
• Cookie security helpers. utils/security_cookies.py sets auth/session cookies with HttpOnly, SameSite, and secure flags (enforced for production or when SECURE_COOKIES=true) and provides symmetric deletion helpers to avoid stale cookies.
• Security headers middleware. SecurityHeadersMiddleware adds CSP, X-Frame-Options (default DENY), X-Content-Type-Options (nosniff), X-Download-Options (noopen), Referrer-Policy, HSTS (when HTTPS is detected), and strips Server/X-Powered-By headers.
• TLS ready. make certs creates local certificates, make serve-ssl runs Gunicorn with TLS, and the client defaults keep SKIP_SSL_VERIFY=false. The container images trust RHEL certificate bundles for outbound TLS.
• Support bundle sanitisation. SupportBundleService redacts passwords, tokens, secrets, and bearer values before writing diagnostic ZIPs. Patterns cover API keys, JWTs, Authorization headers, and database URLs.
• Configuration masking. /admin/config/settings hides sensitive keys using the mask_sensitive helper to prevent secret exfiltration through the Admin UI/API.
• Hardened containers. Containerfile.lite builds on patched RHEL UBI 10 ↦ scratch, installs dependencies in a venv, strips debugging symbols, removes package managers, deletes setuid/setgid binaries, creates a non-root UID 1001, preserves the RPM DB for scanning, and sets security-oriented runtime env vars.
• Logging controls. LoggingService centralises formatting with JSON or text output, supports log rotation, and the support bundle honours size/line caps to avoid log leakage.

Input Validation & Guardrails

• SecurityValidator centralises sanitisation. mcpgateway/validators.py enforces length limits, safe character sets, scheme allowlists, JSON depth (MAX_JSON_DEPTH), and detects dangerous HTML/JS patterns across every Pydantic schema.
• Schema-driven enforcement. mcpgateway/schemas.py applies the validator to tool names, URLs, resource URIs, descriptions, prompts, and JSON payloads, trimming or rejecting unsafe values before they hit business logic.
• Guardrail plugins. Content filters (PII, harmful content, markdown cleanup), output length guards, regex filters, and SQL sanitizers run as pre/post hooks to block malicious prompt/tool/resource usage.
• Rate limiting & quotas. Admin routes use the rate_limit decorator to enforce per-IP quotas (admin.py), bulk import has configurable ceilings (MCPGATEWAY_BULK_IMPORT_RATE_LIMIT, MCPGATEWAY_BULK_IMPORT_MAX_TOOLS), and runtime limits (tool_rate_limit, tool_timeout, tool_concurrent_limit) prevent resource exhaustion.
• Header and payload hygiene. Passthrough headers strip control characters, clamp values to 4 KB, and refuse malformed names; request/response retries honour jitter and backoff to mitigate abuse.
• Secure invitation flows. Team invitations use cryptographically random, URL-safe tokens, enforce owner-only issuance, respect expiry, and prevent over-subscribing teams (TeamInvitationService).

Operational Security & Monitoring

• Startup enforcement. validate_security_configuration() blocks boot when critical issues remain and REQUIRE_STRONG_SECRETS=true, and otherwise prints actionable warnings (default secrets, disabled auth, SSL verification overrides).
• Security event logging. SECURITY_LOGGING_ENABLED persists authentication attempts, authorization failures, and security-relevant events to the security_events table for audit and investigation. SECURITY_LOGGING_LEVEL controls verbosity: all (every event, high DB load), failures_only (default, authentication/authorization failures), or high_severity (critical events only). Disabled by default for performance.
• Continuous telemetry. Permission checks, OAuth flows, and token usage log structured events with timestamps, IP addresses, user-agent strings, span attributes, and success/failure flags for downstream monitoring.
• Security tooling baked into the build. The Makefile exposes make security-all, make security-scan, make security-report, make bandit, make semgrep, make dodgy, make gitleaks, make trivy, make grype-scan, make snyk-all, and make fuzz-security, providing repeatable security automation for CI/CD.
• Observability hooks. OpenTelemetry exports (when configured) tag spans with error flags, latency, and success status, supporting tracing-based detection of anomalies.
• Support bundle hygiene. Operators can gather diagnostics without leaking credentials thanks to sanitisation routines and configurable size/time limits.

Production Hardening Checklist

• Set production posture. Run with ENVIRONMENT=production, configure APP_DOMAIN and explicit ALLOWED_ORIGINS, and leave SKIP_SSL_VERIFY=false.
• Harden secrets. Rotate JWT_SECRET_KEY, AUTH_ENCRYPTION_SECRET, PLATFORM_ADMIN_PASSWORD, and database credentials; enable REQUIRE_STRONG_SECRETS=true so weak values stop startup.
• Keep Basic auth disabled. Leave API_ALLOW_BASIC_AUTH=false (default) and DOCS_ALLOW_BASIC_AUTH=false (default). Use JWT tokens for API access.
• Keep auth mandatory. Maintain AUTH_REQUIRED=true, MCP_CLIENT_AUTH_ENABLED=true, and only enable TRUST_PROXY_AUTH behind a trusted authentication proxy.
• Disable unused surfaces. Leave MCPGATEWAY_UI_ENABLED=false, MCPGATEWAY_ADMIN_API_ENABLED=false, MCPGATEWAY_BULK_IMPORT_ENABLED=false, MCPGATEWAY_A2A_ENABLED=false, and MCPGATEWAY_CATALOG_ENABLED=false unless you actively use them.
• Leave header passthrough off. ENABLE_HEADER_PASSTHROUGH=false (default) should only change after reviewing downstream requirements and allowlists.
• Secure the data plane. Terminate TLS with real certificates (make certs/make serve-ssl or a fronting proxy), and prefer PostgreSQL/MySQL with TLS over SQLite in production.
• Monitor activity. Ship token_usage_logs, email_auth_events, audit trails, and structured logs to your SIEM/observability stack; alert on repeated failures or blocked requests.
• Automate security checks. Integrate the security Make targets into CI/CD so images, dependencies, and IaC are scanned before deployment.

Planned & In-Progress Enhancements (🚧 Planned)

The items below are active roadmap work or design explorations. Track status in docs/docs/architecture/roadmap.md and the linked GitHub issues.

Authentication & Authorization

• 🚧 Attribute-Based Access Control (ABAC) — Use user attributes and resource metadata to supplement RBAC for multi-tenant servers (#706).
• 🚧 Policy-as-Code enforcement — Integrate Rego/OPA policies directly into gateway decisions beyond the current optional plugin (#271).
• 🚧 Per-virtual-server API keys & conditional capabilities — Expand token scopes to cover tool-level capability grants and auto-expiry policies.

Data Protection & Secrets

• 🚧 HashiCorp Vault & external KMS — Native secret backends for tool credentials, JWT keys, and OAuth secrets (#542).
• 🚧 mTLS and certificate pinning — Stronger upstream trust requirements for MCP servers with automatic pin management (#568).
• 🚧 Data Loss Prevention (DLP) — Inline scanning for sensitive payloads with redact-or-drop policies.
• 🚧 Advanced cryptography — Evaluating TEEs, HSM-backed signing, and post-quantum algorithms for long-lived deployments.

Runtime & Infrastructure Security

• 🚧 Container runtime enforcement — Integrations for Falco, AppArmor/SELinux profiles, seccomp, and CapDrop during container deployment.
• 🚧 Service mesh alignment — Dedicated Istio/Linkerd blueprints for mTLS, traffic policies, and zero-trust networking.
• 🚧 Federated attestation — Signing/verification workflow for MCP gateways and servers to establish trust before federation link-up.
• 🚧 Sandboxed execution — gVisor, Firecracker, or WebAssembly sandboxes for untrusted MCP servers and plugins.

Monitoring & Governance

• 🚧 Dynamic security posture scoring — Automated risk evaluations enriched with runtime metrics and audit history.
• 🚧 Behavioral analytics — ML-assisted anomaly detection for unusual tool usage, prompt patterns, or federation activity.
• 🚧 Immutable audit trails — Evaluating tamper-resistant storage (e.g., append-only or ledger-backed logs) for high-assurance environments.

These features remain aspirational until the associated PRs merge. Expect the documentation to move them into the "Current" sections when code lands.

Additional References

• Configuration reference: .env.example and README.md cover every toggle in more depth.
• Security policy: SECURITY.md documents vulnerability disclosure expectations.
• Multi-tenancy details: docs/docs/architecture/multitenancy.md digs deeper into RBAC and team scoping.
• Deployment guidance: docs/docs/deployment/helm.md and Containerfile.lite showcase hardened deployment patterns.